A Claude Code skill that bootstraps a brand-new project with a structured, self-contained knowledge base — then hands you ready-to-run prompts to build the project from that KB.
You answer a guided interview. It writes the docs. You get a spec-first project instead of a blank repo.
/kb-kickoff
The skill runs in three phases:
- Builds the KB — interviews you across functional analysis, technical analysis, infrastructure, existing design, and key decisions, then writes a populated, cross-linked documentation tree.
- Emits implementation prompts — writes
PROMPTS.md, a sequenced set of copy-paste prompts that drive an agent to scaffold and implement the project, each grounded in the specs it depends on. - Stops. It does not write application code unless you explicitly ask. The deliverable is the KB + the prompts.
The doc structure and conventions mirror a real product knowledge base: a functional/technical split, ADRs for decisions, frontmatter on every doc, and [[wikilinks]] between related docs.
A new repo usually starts as a blank folder and a vague idea. Code gets written before anyone writes down what the system is supposed to do, what the rules are, or why a choice was made. kb-kickoff inverts that: it forces the thinking up front into durable, navigable docs, and the implementation prompts keep the build anchored to those docs instead of to guesswork.
A lean, portable, per-project tree (no external vault dependency):
<project>/
CLAUDE.md # how an agent should use this KB
README.md # project summary + index of the KB
00-meta/
taxonomy.md # allowed frontmatter values
templates/ # blank templates so the KB can grow consistently
architecture.md # high-level technical architecture (type: product)
functional/ # one doc per user flow — user POV, no code
technical/ # one spec per feature/area — business rules, edge cases, data model
api/ # endpoint contracts (if there's an API)
infrastructure/ # topology, deployment, security
decisions/ # one ADR per significant choice
design/ # points to existing design (prototype/Figma) — read, don't invent
PROMPTS.md # sequenced implementation prompts (the build plan)
Empty branches are omitted gracefully (no API → no technical/api/).
- Frontmatter on every doc (
title, id, type, status, owner, created, updated, review_by, audience, sensitivity, tags, related, ai_usage) plus type-specific keys. - Functional ≠ technical. Functional docs are user-POV only — no endpoints, schemas, or code paths. Technical specs carry verifiable business rules, edge cases, states/transitions, and a
code_referencesarray to fill once code exists. - Everything ships as
status: draftuntil a human reviews it. > TODO:beats fiction. Unknowns are flagged, never fabricated.- You pick the language at the start; all docs are written in it (frontmatter keys stay English).
/kb-kickoff # interview + generate KB in ./<project-slug>/
/kb-kickoff <project-name> # pre-seed the project name
/kb-kickoff <path> # generate at a specific path
/kb-kickoff --from <brief-file.md> # read an existing brief first, ask only what's missing
/kb-kickoff --design <dir-or-file> # design already exists here: read it, don't invent it
/kb-kickoff --resume # continue an interrupted interview from what's on disk
This is a Claude Code skill. Drop the folder into your skills directory:
git clone https://github.com/nzereik/kb-kickoff.git ~/.claude/skills/kb-kickoffThen, optionally, register the trigger in your ~/.claude/CLAUDE.md so it's always discoverable:
# kb-kickoff
- **kb-kickoff** (`~/.claude/skills/kb-kickoff/SKILL.md`) - bootstrap a new project: guided interview → structured knowledge base → implementation prompts. Trigger: `/kb-kickoff`
When the user types `/kb-kickoff`, invoke the Skill tool with `skill: "kb-kickoff"` before doing anything else.Restart Claude Code (or start a new session) and run /kb-kickoff.
kb-kickoff/
SKILL.md # the skill definition (what Claude Code loads)
reference/
templates-skeleton.md # fallback doc templates, keeps the skill self-contained
README.md # this file
MIT — see LICENSE.